andygrove opened a new issue, #5848:
URL: https://github.com/apache/datafusion-comet/issues/5848
### Describe the bug
`CometCodegenSuite."codegen dispatch coverage survives the decimal promotion
rewrite"` fails on main, taking down the `[expressions]` job on all five Spark
profiles (3.4, 3.5, 4.0, 4.1, 4.2) for every PR.
```
- codegen dispatch coverage survives the decimal promotion rewrite ***
FAILED ***
Some(Set("hypot", "cast", "checkoverflow", "add")) did not contain
Set("hypot")
(CometCodegenSuite.scala:431)
```
The assertion is stale rather than the behaviour being wrong. #5610
(`f4b8efdd4`) deliberately started recording every expression in a dispatched
subtree, not just its root: the whole subtree is bound and closure-serialized
into one kernel, so all of it ran in the JVM. Its own commit message says
"Recording the descendants of a dispatched subtree changes q78's dispatcher
list and its codegen-dispatch count", and it regenerated the q78 plan-stability
goldens for exactly that. This assertion, which predates the change and expects
the promoted `Hypot` subtree to contribute `"hypot"` alone, was not updated
with them.
Identical failure and identical actual set on all five profiles, so it is
deterministic and not version-specific.
### Steps to reproduce
On a clean checkout of main:
```
./mvnw test -Dsuites="org.apache.comet.CometCodegenSuite codegen dispatch
coverage survives the decimal promotion rewrite" -Dtest=none
```
### Expected behavior
The lifted set names the whole dispatched subtree: `hypot`, the `cast`, the
promotion's `checkoverflow` wrapper and the decimal `add`.
### Additional context
First seen in CI on an unrelated PR (#5840). Nothing has landed on main
since #5610 that touches the coverage lift.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
